Object detection - Dog tagged as Person

[JCAK]

I have it set to "medium" but I did the test in both modes, High and Medium, and uploaded the videos so you can see the result.

https://imgur.com/a/2ZgpVbz

[YrbkMgr]

My first thought was... man, you can't solve this in two days... and online? Then it occurred to me that you're using a trial version - I have no idea what if any effect that may have.

I have never seen DS work out of the box. It requires fine tuning. It has taken months for me to get DS successful. From an expectations standpoint, that's something to consider. The overview of what you have to do is: Configure camera's encoder streams correctly - first step. Then you have to analyze the DAT files produced by DeepStack. That will help to pinpoint where to focus attention. It's not a trivial process, although there are resources and friendly help on how to do that. Point is though, if you're in a hurry so you can make a decision in the next two days, you would be better off hinging your decision on some other criteria.

In my opinion, there isn't any better VMS software (non-enterprise) than Blue Iris. Even without DeepStack, the robust and extensive features in v5 are a bargain. If you buy BI and later regret it, how bad did you get burned? 60 bux. I say, do like everyone else, pull the trigger, and learn how to use the features at your leisure.

[JCAK]

The only limitation about demo version is the days of trial (15) and BI don't speak nothing about restrictions from any other kind. I understand your point of view about burning 60 bux, but I prefer to buy something that works well. This forum have a lot of people asking about issues after upgradeing versions or fails after add a camera, things that work and suddenly stop working, etc, etc. and blaming BI because they say that the same thing always happens, something works and something else stops working. Whay buy a lotery? I could buy it but at least if it works what I need.
Yes, BI look like a swiss knife but the AI feature is essential for me because I have a NVR without IA and whant move to a more professional solution. I need process outdoor images with moving tree leaves, dogs runnings or birds, bees and spiders cross the camera lens. I need a IA capable to filter all this and trigger an alert only when a person enters the field. I spent a lot of time testing BI and is great in many ways but I need solve what I tell you or BI is not for me.

I have no problem not detecting a dog, but at least not detecting it as a person. Look at this other example in attached file:

[YrbkMgr]

I get what you're saying. Totally. I have been using BI since v3, and didn't go from 4->5 until a year after its release. Been on a steep learning curve to figure it out. BI Support used my support ticket to demonstrate the use of the DeepStack tab in fine tuning. There is no way that BI with DS cannot do what you, and thousands of others want to do. Your application is the poster child for success. That's my personal opinion.

If you want to fine tune it, you have to put the work in just like everyone else. It'll take time, but help is here. So, well... yeah, that.

[JCAK]

Last night I was working on this matter. I found the problem and no amount of tweaking I make to BI is going to solve it. The recognition problem is because BI is using the models that come by default with DeepStack and they have not created a better model. I was searching the internet and found a couple of custom models for DeepStack, tried them out and saw the difference, for example:

https://github.com/MikeLud/DeepStack-Se ... era-Models

From the above link I downloaded the file "general.pt" and instead of using "Default object detection" I changed it to "Use custom model folder". This pre-trined model file only recognise 6 objects instead the default 80 objects that come with DeepStack, this custom file is a truncated version from the original DS file. The performance difference was amazing, the speed of processing for object detection was 3x o 4x. I could see how the orange frame is updated in less time, it appears many more times showing and updating the confidence level.

What I mean by all this is that Blue Iris should provide the solution for better detection by creating its own pre-trained model file. Most people don't have the knowledge or the time to create these types of files. It is not a question of spending 60 dollars like that, what one expects is something else.

[YrbkMgr]

I found the problem and no amount of tweaking I make to BI is going to solve it. The recognition problem is because BI is using the models that come by default with DeepStack and they have not created a better model.

Respectfully, I do not agree with either of those statements. I'm reminded of a quote I heard "The soundest reasoning leads to the wrongest conclusions when the premises are false". You're seeing the Sun come up in the East when the rooster crows and concluding that the rooster causes sunrise. If your conclusions were true, they would be true for everyone. And they're not. Please know, I am not trying to convince you of anything - I am only sharing what I know to be true - do with that what you will.

The only way to solve the problem is by analyzing what is happening. I'm willing to help you do that. If you are of the opinion that for whatever reason, I am not credible or don't understand your problem, I'll stop trying to help. You get to say.

But know this... in an orchestra, if every instrument is not tuned, the performance sounds like crap. No different here.

[JCAK]

Hello,

I'm here again, with good and bad news.
The good news is that after today we have all the time we want to test object detection, so buy Blue Iris
The bad news is that I'm still having problems and now I spent 60 bucks

I send you two new images from today. All the settings remain the same as the ones you indicated. I don't know what else to do so that it doesn't detect my dog as a person. This happens with all cameras.

Could you help me to fix this issue, please?

[MikeBwca]

It's not that your dog is being detected as a person... your 'person' model is detecting your dog! Weird. You don't live in Transylvania?

- Increase the confidence level of the 'person' label. This can be done specifically, separate from the others, by suffixing any label with ':##'. For example 'person:90'. Start high, walk around, let your dog run around. Then lower until 'person' detects your dog, and raise it up a nitch.

- Your using models. See page 74. you can include/exclude various models.

- You may want to force the issue by training a new model. Search 'Deepstack train model'.

[JCAK]

Increasing to a very high confidence level is not what I want, because it has detected the dog with 81%.

Using other models is an option I've suggested before, but someone in the group told me that that probably wasn't the problem and that I should tweak more things. I don't know what else to adjust and I think the problem is the models that come with DeepStack. I wanted to avoid the creation of new models, honestly I would like Blue Iris to have its own and more complete library than the one that comes with DeepStack, but if there is no other option I will have to do it.

I have found several posts from other users who complain about the same thing, about the lack of accuracy in object detection. It is a real problem when used in places where there is a lot of wildlife or pets. One suggestion I make is that Blue Iris could create a library of various detection models, so that the user can combine them according to the environment in which they are going to use it.

[YrbkMgr]

Increase the confidence level of the 'person' label. This can be done specifically, separate from the others, by suffixing any label with ':##'. For example 'person:90'. Start high, walk around, let your dog run around. Then lower until 'person' detects your dog, and raise it up a nitch.

I had no idea about the label confidence override - nifty tip. Looking into it, it seems that the feature is used to override the confidence percent when there is more than one label in the confirm box. If you only have one label (person), there's no need or benefit to overriding it. Check my homework MikeBwca... thoughts?